Recurrent 3D Pose Sequence Machines

摘要: 3D human articulated pose recovery from monocular image sequences is very challenging due to the diverse appearances, viewpoints, occlusions, and also inherently ambiguous imagery. It thus critical exploit rich spatial temporal long-range dependencies among body joints for accurate sequence prediction. Existing approaches usually manually design some elaborate prior terms kinematic constraints capturing structures, which are often insufficient all intrinsic structures not scalable scenarios. In contrast, this paper presents a Recurrent Pose Sequence Machine(RPSM) automatically learn image-dependent structural constraint sequence-dependent context by using multi-stage sequential refinement. At each stage, our RPSM composed of three modules predict based on previously learned 2D representations poses: (i) module extracting representations, (ii) recurrent regressing poses (iii) feature adaption serving as bridge between enable representation transformation domain. These then assembled into prediction framework refine predicted with multiple stages. Extensive evaluations Human3.6M dataset HumanEva-I show that outperforms state-of-the-art estimation.